Imagined Rollouts are Kinematic, Not Dynamic: A Diagnosis of Long-Horizon World-Model Failure
Abstract
World models exhibit long-horizon failures due to kinematic rather than dynamic imagination, as demonstrated by measuring imagined kinematic-consistency error which remains flat while policy rewards collapse across friction boundaries.
Long-horizon failure in world models is conventionally attributed to compounding error, a generic framing that does not distinguish what kind of error compounds. We propose a kinematic-vs-dynamic reframing: world models tend to imagine kinematically rather than dynamically. We operationalize this as the imagined Kinematic-Consistency Error, a per-step diagnostic that measures how far a rollout departs from a closed-form kinematic null, paired with a perturbation protocol that tests whether iKCE responds when physical conditions cross a regime boundary. We instantiate the diagnostic on a released DreamerV3 checkpoint trained on DMC walker-walk, where imagined iKCE runs roughly two orders of magnitude above that of matched real-physics rollouts. Across a friction sweep that crosses the gait-collapse boundary, the model's iKCE stays statistically flat even as the trained policy's reward collapses through the same range, providing the kinematic-not-dynamic signature. The diagnostic distinguishes kinematic from dynamic imagination at horizons longer than the embodiment's gait period.
Community
TL;DR: We introduce iKCE (imagined Kinematic-Consistency Error), a cheap,
falsifiable test-time diagnostic that sharpens the usual "compounding error" story into
a concrete question: what kind of error compounds in a world model's long-horizon
imagination, kinematic or dynamic?
On a released DreamerV3 walker-walk checkpoint, imagined rollouts sit ~180× above a kinematic null yet stay statistically flat across a friction sweep that crosses the gait-collapse boundary the real policy falls off of. The model imagines perfect walking while staying blind to the friction that decides real-world success.
🏆 Accepted at the Robotic World Models Workshop @ RSS 2026
Why it matters: "Long-horizon world-model failure" is usually hand-waved as
compounding error. We give a mechanistic reframing (kinematic vs. dynamic) plus a
one-line test any world model with a decodable kinematic state can be run through.
Key highlights:
- A signature, not a score. The diagnostic is regime-invariance, flat iKCE
(slope CI contains 0) across a physical-regime boundary, not absolute magnitude.
A trivially kinematic predictor scores zero iKCE, so low iKCE ≠ dynamic imagination. - Controls carry the claim. Three seeds, an actor-training-horizon ablation, a
domain-randomization checkpoint (WM in-distribution at every µ), and a joint-noise
positive control showing the WM does respond to kinematic perturbations, just not
dynamic ones. - Open source: code, checkpoints, and perturbation-sweep CSVs to reproduce every figure.
Happy to discuss the framing, where the kinematic/dynamic line gets fuzzy, and whether the signature survives on richer contact dynamics (quadruped/humanoid, or driving WMs with an ego-pose head).
Models citing this paper 0
No model linking this paper
Datasets citing this paper 0
No dataset linking this paper
Spaces citing this paper 0
No Space linking this paper